LocalRecall MCP server provides 11 well-structured tools with consistent naming (verb_noun pattern: search, add_document, list_files, delete_entry, etc.) and mostly clear descriptions. All tools are explicitly registered in pkg/toolset/localrecall/toolset.go with JSON Schema input definitions. However, several quality gaps reduce the score: (1) Output schemas are undocumented, no tool describes what fields it returns, which breaks the pattern of documented responses; (2) Parameter descriptions, while present, are often minimal (10-30 chars) and lack guidance on constraints, formats, or dependency relationships; (3) No error handling guidance, tools do not explain what to do if a search fails, a collection doesn't exist, or a deletion succeeds; (4) Destructive operations (delete_entry, reset_collection) lack confirmation/dry-run patterns; (5) Missing parameter constraints (e.g., max_results has no min/max bounds, min_similarity lacks validation rules); (6) No discussion of mutual exclusivity in add_document (file_path vs file_content) beyond a bare comment. The server is well above baseline (typical STDIO servers ~45-50) due to solid naming and schema presence, but falls short of A-grade (80+) due to incomplete output documentation and minimal parameter guidance.

• Document output schema for every tool. Example for search: 'Returns an array of matched documents: [{id: string, filename: string, content: string, similarity: number, metadata: object}]. Results are sorted by similarity (highest first).'
• Expand parameter descriptions to 50 - 100 chars with explicit constraints. Example: 'Maximum number of results to return (default: 5, range: 1 - 100). Larger values increase latency and token usage.'
• Add error handling guidance to tool descriptions. Example for search: 'If no results match, consider lowering min_similarity or broadening the query. If the collection doesn't exist, call create_collection() first.'
• For add_document, state mutual exclusivity in the InputSchema description: 'Provide exactly one of file_path (upload from filesystem) or file_content (inline string). Both cannot be specified together.'
• Add minValue/maxValue constraints to numeric parameters in the JSON Schema. Example: 'max_results': {type: 'number', minimum: 1, maximum: 100, description: '...'}
• Implement confirmation/dry-run for destructive tools. Add a 'confirm_delete' boolean parameter to delete_entry and reset_collection. If false, return 'Dry run: would delete X items. Call again with confirm_delete=true to proceed.'
• For tools that require a collection_name, note if it's auto-resolved from a default or must be explicitly provided. Currently, buildCollectionTool() handles this, but the tool description should clarify the behavior to the LLM.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry